Some geological configurations, like sedimentary basins, are prone to site effects. Basins are often composed of different geological layers whose properties are generally considered as spatially homogeneous or smoothly varying. In this study, we address the influence of small-scale velocity fluctuations on seismic response. For this purpose, we use the spectral element method to model the 2-D SH wave propagation on a basin of 1.1 km long and ≈60 m deep, representing a 2-D profile in the city of Nice, France. The velocity fluctuations are modelled statistically as a random process characterized by a Von Karman autocorrelation function and are superimposed to the deterministic model. We assess the influence of the amplitude and correlation length of the random velocities on the surface ground motion. We vary the autocorrelation function’s parameters and compute seismic wavefields in 10 random realizations of the stochastic models. The analyses of our results focus on the envelope and phase differences between the waveforms computed in the random and deterministic models; on the variability of ground motion intensity measures, such as the peak ground velocity, the pseudo-spectral acceleration response; and the 2-D basin response (transfer function).We find that the amplitude of fluctuations has a greater effect on the ground motion variability than the correlation length. Depending on the random medium realization, the ground motion in one stochastic model can be locally amplified or deamplified with respect to the reference model due to the presence of high or low velocity contrasts, respectively. When computing the mean amplification of different random realizations, the results may be smaller than those of the reference media due to the smoothing effect of the average. This study highlights the importance of knowing the site properties at different scales, particularly at small scales, for proper seismic hazard assessment.